Both your colleagues were right.
The first one was referring to substantive significance, and was right because
the choice of a reference category is absolutely irrelevant: your substantive results
would be the same, whatever reference category you choose; the choice is like
choosing a unit of measurement for the odds ratios of other categories.
Your other colleague was also
right. He/she was concerned with the statistical significance of the estimate:
if ANY coefficient is estimated on the basis of very few cases its estimated
value would be unreliable, and when the estimate plays the role of reference
category the lack of stability will be transmitted to all odds ratios in which
that ref category is involved.
Hector
De: SPSSX(r)
Discussion [mailto:[hidden email]] En nombre de Claus D.
Hansen
Enviado el: Monday, October 25, 2010 4:23 PM
Para: [hidden email]
Asunto: Choice of reference category
Dear members of the list,
I’m been discussing some points about choice of
reference category in logistic and cox regressions with a colleague.
I’ve read several places that the choice of reference
category in principle is arbitrary and of no importance to the results of your
analyses. Others state that choosing a reference category with too few cases
can lead to unstable and non-robust estimates. Can anyone explain this to me
and perhaps even better point me to a reference or two where this matter is
discussed in more detail perhaps even using an example to illustrate it.
Thank you in advance,
Claus D. Hansen
Aalborg
Se certificó
que el correo entrante no contiene virus.
Comprobada por AVG - www.avg.es
Versión: 8.5.448 / Base de datos de virus: 271.1.1/3206 - Fecha de la versión:
10/24/10 06:34:00
Locked
